A Minimalistic Approach to Tactile Sensing: Object Classification Using Vibration Signals

Abstract

This paper presents a novel approach to object classification using tactile vibrations. While work in robotic perception is becoming increasingly multi-modal, the problem of object classification remains predominantly approached in the visual domain. We challenge this standard by showing that the modality of touch is capable of performing object classification to a high degree of accuracy. We contribute to the growing literature on haptic perception by introducing a new concept of tool-mediated object classification using a con- tainer to enclose the objects and a simple sensor. In this study, a single vibration sensor is used to detect the vibrations of the container induced by shaking objects in the container. Multiple supervised learning classifiers are compared and used to train models on the ex- tracted features of the vibrations. Here, we report that a contact microphone paired with a glass container and a linear SVM classifier yields an average pairwise classification of 99%. Furthermore, experimental results show that a 98% accuracy for multi-class classi cation across ten different objects is achieved using a plastic cup, contact microphone, and a Sparse Representation-based Classifier.